Evolutionary construction of a simulator for real robots

In order to acquire useful motions of a real-world robot, it is necessary to carry out learning in a real environment. However, learning is difficult within a real environment. In addition, the acceleration of learning is required for a practical execution. We propose an approach to the learning acceleration using data retrieved from the real environment. This consists of the method of automatically constructing the simulator from real data and of learning a robot controller with the simulator. The experimental results suggest that our GP-based technique enables the effective controller learning.